Andy Kessler: how the AI data-centre boom ends (opens the source in a new tab)
- The WSJ columnist argues that today’s brute-force AI, giant models on costly chips in power-hungry data centres, will go the way of mainframes: still sold, but no longer where the growth is.
- He says the “factor of 10” shift has begun: comparable Chinese open-weight models charge about 90% less per token than leading US labs, and compressed models now run on PCs and phones with no token costs.
- If prices fall 90%, demand must rise 900% just to stand still, so he warns of data-centre halts and debt defaults, and asks whether 2027 is peak data centre.
- ~90%
- cheaper tokens, Chinese models
- 900%
- demand rise needed to stay flat
Why it matters Data-centre spending is financed on today’s token prices; if prices collapse faster than demand grows, lenders feel it before chipmakers do.
Other side Cheaper tokens can unlock new uses, and Kessler himself says training, science and agents still need big chips, so the build-out could slow without collapsing.